More The Changelog: Software Development, Open Source episodes

We now have Kerrs Law (Interview) thumbnail

We now have Kerrs Law (Interview)

Published 4 Nov 2026

Duration: 2:27:04

Engineering bottlenecks stem from unsynchronized mental models, requiring deliberate review processes and simple Unix-inspired primitives to manage agentic workloads effectively.

Episode Description

Dave Kerr created hacker laws, and now he finds himself reviewing agent-written code for them. Back on the show after six years, Dave names what he se...

Overview

Dave Kerr, creator of the Hacker Laws open-source repository and author of Effective Shell, discusses how AI coding agents have shifted the engineering bottleneck from writing code to maintaining shared mental models across a team. He argues that the limiting factor to build and scale is "how well you keep your team's mental models in sync with reality," and deliberately slows his team down with spec reviews and code reviews not for quality control but so that when a bug arrives, someone can reason about the system without starting from first principles. His team builds agentic workloads on Kubernetes using a tool called Arc, defining agents as YAML resources with model references, system prompts, tool associations, and permissions, treating them like microservices within existing SRE tooling.

Kerr reflects on how the Unix philosophy of small, composable tools has been vindicated by coding agents that stitch together bash, grep, and ls into complex pipelines. He describes grounding agents by requiring three idiomatic examples before they invent new patterns, reducing cognitive burden on developers who already know tools like kubectl or the GitHub CLI. He also recounts a security audit of a hobby project where the LLM produced a "fundamental mess" of symmetric encryption because he never explicitly asked for security, illustrating the need to direct the model toward non-functional requirements.

The conversation touches on open-source grief - watching handcrafted side projects compete with one-shots - and a provocative question about whether frontier labs might one day sell placement in architectural recommendations, analogous to how Google monetizes search rankings. Kerr also shares a conference talk on how AI can dysregulate neurodivergent developers, feeding grandiosity and maladaptive creativity, and the importance of boundaries like shutting down creative work after six p.m.

What If

  • What if you required three idiomatic examples before letting an agent write any new code?

    • Move: Before accepting agent-generated code, prompt it to produce three real-world idioms (from kubectl, the GitHub CLI, or your existing codebase) that demonstrate the pattern it is about to implement. Reject any output that invents a novel approach when a standard one exists.
    • Why Now?: Kerr's team found agents would "invent something" until they were "forcibly grounding" them with this rule, and LLMs already know common flags like --dry-run from training data, so leveraging idioms yields better results at zero extra cost.
    • Expected Upside: A proposal - reduced cognitive burden for future reviewers, fewer principle-of-least-surprise violations, and less time spent debugging clever-but-unnecessary abstractions.
  • What if you made every agent-produced pull request ship with a recorded terminal session or screenshot instead of just passing tests?

    • Move: Instruct your coding agent to run the CLI, capture a video or SVG rendering of the output, and attach it to the PR. Use a tool like Show Right or even asciinema to generate the artifact automatically.
    • Why Now?: Kerr notes that "unit tests passing" is "hygiene, not evidence," and that reviewing a 30-second recording of htop or canines running lets you "eyeball" correctness without reading code, saving hours of manual verification.
    • Expected Upside: A possibility - faster review cycles, earlier detection of "snafus that will catch us later," and a growing library of visual regression artifacts you can diff across releases.
  • What if you explicitly told the agent your security requirements before it wrote a single line, even for a hobby project?

    • Move: Add a standing instruction or CLAUDE.md rule: "Assume any code touching passwords, sessions, or external input must use asymmetric encryption, signed certificates, and a threat-model comment block. If the task does not mention security, ask before proceeding."
    • Why Now?: Kerr's markdown-sharing hobby project produced symmetric-encryption passwords "passed around in silly ways" because he never stated the security need; the LLM was "perfectly good" at the task he described, and the gap was entirely in the prompt.
    • Expected Upside: A proposal - avoiding the "genie in a bottle" trap where the model gives you something that works but "would never pass any kind of pen test," and eliminating the costly retrofit of asymmetric crypto after the fact.

Takeaway

  • Add a mandatory spec-review pass before implementation: assign one person to review the specification and a different person to review the implementation, creating a deliberate choke point that keeps mental models synchronized without slowing the agent's output.
  • Before accepting any agent-generated code, require it to cite three existing idioms (from your codebase, kubectl, the GitHub CLI, or similar) that validate the pattern; reject novel approaches where a standard one exists.
  • Instruct your agent to produce a recorded terminal session, screenshot, or SVG rendering of the CLI output as a PR artifact, treating that evidence - not just passing unit tests - as the gate for review.
  • Write a standing security instruction into your agent configuration (CLAUDE.md, system prompt, or equivalent) that mandates asymmetric encryption, signed certificates, and a threat-model comment for any code touching credentials or external input, even in hobby projects.
  • When the agent proposes building something, pause and ask: "Is this a solved problem with three or four existing solutions?" Default to buy (or open-source) before build, especially for auth, file storage, and caching, to avoid the model's "down to build" bias.

Recent Episodes of The Changelog: Software Development, Open Source

22 Oct 2026 Building what's next (Interview)

Resend optimizes its email infrastructure for AI agents by implementing markdown pricing pages, RFC-compliant headers, and granular permissions to support three million users.

9 Oct 2026 From zero to NanoClaw (Interview)

"AI agents are transforming software development and business operations by automating tasks, rewriting code, and disrupting workflows, while raising challenges like license enforcement, governance, and the future of knowledge work in an AI-driven world."

3 Sept 2026 Forking Cal.com to closed source (Interview)

"AI-generated code floods open source projects, obscuring vulnerabilities and overwhelming maintainers, prompting shifts to private repositories and cloud-based development as security risks escalate."

25 Aug 2026 Postgres at PlanetScale (Interview)

"PlanetScale launched a $5 PostgreSQL plan, detailed metal deployment for performance, previewed sharded PostgreSQL product Niki, and discussed AI workloads, CI/CD challenges, and engineering-driven scaling strategies."

21 Jul 2026 Canary tokens and digital tripwires (Interview)

"Thinkst, a 50-person cybersecurity firm, specializes in Canary honeypots and Canary Tokens, offering simple, no-maintenance detection tools with $22.5M ARR, no price hikes in a decade, and AI-driven threat awareness."

More The Changelog: Software Development, Open Source episodes